Skip to content

feat(setup): wizard gates on prereqs, probe defaults to No, setup --smoke - #39

Merged
giveen merged 1 commit into
mainfrom
feat/wizard-gates-and-smoke
Oct 2, 2026
Merged

giveen merged 1 commit into
mainfrom
feat/wizard-gates-and-smoke

Conversation

@giveen

@giveen giveen commented Oct 2, 2026

Copy link
Copy Markdown
Owner

Release-readiness: a fresh user could burn 30 minutes before discovering a missing tool or dead model. Fail fast instead.

Changes

  • Wizard step 0 — prereq gate: runs setup::check_all() first. Missing Required tools bail immediately with fix hints (run kernelopt setup table); missing Recommended only warns (non---yes mode). --yes mode bails the same way.
  • prompt_llm defaults to No on unhealthy probe: "Use it?" defaults Yes only when the model is listed AND the tool call works. Unlisted-but-working models get an explicit warning; broken models require an explicit y. Closes the fledge-alpha-free footgun (dead model accepted by pressing Enter).
  • kernelopt setup --smoke: compiles a minimal __global__ kernel with nvcc (--ptx) or hipcc (--genco) to prove compiler + headers work — independent of --install, no GPU needed. Verified live: smoke: nvcc compiled a minimal kernel (headers + driver toolchain OK).

Validation

  • cargo build --release clean; cargo test 122 passed; pytest 21 passed.
  • setup --smoke verified live on this box.
  • Wizard gate + probe logic by inspection (interactive paths); gate only triggers on missing Required, all present here.

Notes

  • Smoke covers compile only, not device execution or the LLM path (those stay in the wizard probe + first run). A device-execution smoke would need a runnable kernel + GPU time; follow-up if wanted.
  • Container image (collapsing CUDA/toolchain install into docker run) remains the bigger fix for broad-audience onboarding; this PR makes bare-metal fail fast instead.

…moke

Release-readiness: a fresh user could burn 30 minutes before discovering a
missing tool or dead model. Fail fast instead.

- Wizard step 0 runs setup::check_all(): missing Required tools bail with
  fix hints before anything else; missing Recommended only warns.
- prompt_llm: "Use it?" now defaults to No unless the probe is fully healthy
  (model listed + tool call works). A broken model is never an Enter key away
  (the fledge-alpha-free footgun); unlisted-but-working models warn explicitly.
- `kernelopt setup --smoke`: compiles a minimal CUDA kernel with nvcc (or
  hipcc) to prove the compiler + headers work, independent of --install.
@giveen
giveen merged commit 4102165 into main Oct 2, 2026
2 checks passed
@giveen
giveen deleted the feat/wizard-gates-and-smoke branch October 2, 2026 13:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant